SafetyOnline
AdvBench
Measures whether adversarial suffixes induce aligned language models to produce targeted harmful behavior.
- Task count
- 520
Safety benchmarks
There are 25 safety in the index.
Results
25 benchmarks
SafetyOnline
AgentDojo
Evaluates agent task utility and resistance to prompt injection while agents use tools over untrusted data.
- Task count
- 97
SafetyOnline
BBQ
Measures social bias in question answering across ambiguous and disambiguated contexts covering nine protected dimensions.
SafetyOnline
BBQ-Lite
Provides a compact BIG-bench version of BBQ for exact-match evaluation of bias-sensitive question answering.
SafetyOnline
BeaverTails
Supports safety classification and preference evaluation with human-labeled question-answer pairs across harm categories.
- Task count
- 330,000
SafetyOnline
BIPIA
Evaluates model and defense robustness against indirect prompt injection across five application tasks.
- Task count
- 5
SafetyOnline
BOLD
Measures social biases in open-ended English generation from prompts spanning profession, gender, race, religion, and politics.
- Task count
- 23,679
SafetyOnline
CrowS-Pairs
Measures stereotypical preferences in masked language models using paired sentences across nine types of social bias.
- Task count
- 1,508
SafetyOnline
CyberSecEval
Evaluates cybersecurity risks from model-generated code and responses to requests that could enable cyberattacks.
SafetyOnline
HarmBench
Evaluates automated red-teaming attacks and model refusal defenses across standardized harmful behaviors.
- Task count
- 510
SafetyOnline
InjecAgent
Evaluates whether tool-using agents follow indirect prompt injections embedded in external tool responses.
- Task count
- 1,054
SafetyOnline
JailbreakBench
Compares jailbreak attacks and defenses on a standardized set of misuse behaviors and evaluation rules.
- Task count
- 100
SafetyOnline
RealToxicityPrompts
Measures whether language models continue naturally occurring web prompts with toxic text.
SafetyOnline
SafeBench
Evaluates autonomous-driving agents in safety-critical simulated scenarios, including adversarially generated cases.
- Task count
- 235
SafetyOnline
SafetyBench
Evaluates large-language-model safety knowledge with Chinese and English multiple-choice questions across seven risk categories.
- Task count
- 11,435
SafetyOnline
SORRY-Bench
Measures the safety-refusal behavior of language models across a balanced taxonomy of unsafe requests and prompt mutations.
- Task count
- 450
SafetyOnline
StereoSet
Measures stereotypical bias and language-modeling ability in sentence-level and discourse-level contexts.
SafetyOnline
StrongREJECT
Evaluates jailbreak responses with a forbidden-prompt dataset and an automated judge designed to penalize empty refusals and ineffective attacks.
- Task count
- 313
SafetyOnline
ToxiGen
Tests hate-speech classifiers on machine-generated implicit toxic and benign statements about 13 minority groups.
SafetyOnline
WinoBias
Measures gender bias in coreference resolution by changing pronoun gender while preserving sentence structure.
SafetyOnline
XSTest
Tests whether safety-tuned language models refuse harmless prompts that superficially resemble unsafe requests.
- Task count
- 450
SafetyOnline
Benchmark MCP Safety
Tests resistance to malicious tool descriptions, tool shadowing, and excess privilege.
SafetyOnline
Benchmark Sycophancy
Tests whether a model resists endorsing a user's increasingly harmful mistaken plan.
SafetyOnline
Do-Not-Answer
Evaluates whether language models decline risky instructions while still answering harmless requests.
- Task count
- 939
Other categories in Online
More benchmarks
Questions
What other kinds of benchmarks are in Online?
Other categories in Online: Multimodal (51 benchmarks), Reasoning (50 benchmarks), Agents (31 benchmarks), Science (24 benchmarks), and Coding (17 benchmarks).
Published by AI Benchmark Index. How we check each benchmark.